arXiv subjects

Zhouzi Li

Publications and source records attributed to Zhouzi Li.

8 recordsLinked to original sources

arXiv2026

SPLIT: SymPathy for Large jobs Improves Tail latency

We study the asymptotic response time tail in the M/G/n multi-server queue with heavy-tailed (regularly varying) job sizes, a setting representative of modern computing workloads. For single-server systems, tail optimization is well understood: under heavy-tailed job sizes, policies such as SRPT that strictly prioritize short jobs are strongly tail optimal, and giving any priority to large jobs is harmful. For multi-server systems, the question has been almost entirely open. This paper gives the first strongly tail-optimal scheduling policies for the M/G/n queue with heavy-tailed job sizes. Our central finding is that the multi-server case is intrinsically different from the single-server case: giving a small amount of ``sympathy'' to large jobs is essential for strong tail optimality. We establish strong (or arbitrarily close to strong) tail optimality across the full stability region, both with and without knowledge of job sizes.

cs.PF↗

arXiv2026

BOA Constrictor: Squeezing Performance out of GPUs in the Cloud via Budget-Optimal Allocation

The past decade has seen a dramatic increase in demand for GPUs to train Machine Learning (ML) models. Because it is prohibitively expensive for most organizations to build and maintain a large GPU cluster, organizations instead choose to rent GPUs from cloud providers. The customer is responsible for devising a policy for (i) deciding how many GPUs to rent at every moment in time to process a stream of ML training jobs and (ii) allocating the rented GPUs among the currently active jobs in the system. Because ML training jobs can be parallelized across different numbers of GPUs, the customer generally has many options for how many GPUs to use for each job. Allocating more GPUs to a single training job will cause the job to complete more quickly. However, the customer pays for each GPU-hour they use, and a training job receives a diminishing marginal benefit from running on additional GPUs. Hence, allocating too many GPUs to a single training job can dramatically increase the overall cost that the customer pays to the cloud provider. This gives rise to a cost-performance tradeoff that customers must balance when running training jobs in the cloud. To balance the cost-performance tradeoff, we develop BOA Constrictor, a new scheduler for ML training jobs which uses a Budget-Optimal Allocation (BOA) policy to squeeze the highest level of performance out of a cloud-deployed GPU cluster given a fixed budget constraint. We explicitly formulate the problem as a budget-constrained scheduling problem and derive the BOA policy which minimizes the average job completion time (JCT) of a stream of arriving jobs subject to the user's budget. For a given budget level, we demonstrate that BOA Constrictor can reduce average JCT by 1.6 times in small-scale implementation experiments and by 2 times in detailed, large-scale simulations compared to state-of-the-art heuristic based schedulers.

cs.DC↗

arXiv2026

Mean field optimal Core Allocation across Malleable jobs

Modern data centers and cloud computing clusters are increasingly running workloads composed of malleable jobs. A malleable job can be parallelized across any number of cores, yet the job typically exhibits diminishing marginal returns for each additional core on which it runs. This can be seen in the concavity of a job's speedup function, which describes the job's processing speed as a function of the number of cores on which it runs. Given the prevalence of malleable jobs, several theoretical works have posed the problem of how to allocate a fixed number of cores across a stream of arriving malleable jobs so as to minimize the mean response time across jobs. We refer to this as the Core Allocation to Malleable jobs (CAM) problem. We solve the CAM problem under a highly general setting, allowing for multiple job classes, each with an arbitrary concave speedup function and holding costs (weight). Furthermore, we allow for generally distributed inter-arrival times and job sizes. We analyze the CAM problem in the mean field asymptotic regime and derive two distinct mean field optimal policies, FW-CAM and WHAM. FW-CAM is interesting because it demonstrates a new intuition: in the mean field regime, job sizes are not relevant in finding an optimal policy. WHAM (Whittle Allocation for Malleable jobs) is interesting because it is asymptotically optimal and also serves as a good heuristic even outside of the asymptotic regime. Notably, none of the policies previously proposed in the literature are mean field optimal when jobs may follow different speedup functions.

cs.DC↗

arXiv2026

LookAhead: The Optimal Non-decreasing Index Policy for a Time-Varying Holding Cost problem

In practice, the cost of delaying a job can grow as the job waits. Such behavior is modeled by the Time-Varying Holding Cost (TVHC) problem, where each job's instantaneous holding cost increases with its current age (a job's age is the time since it arrived). The goal of the TVHC problem is to find a scheduling policy that minimizes the time-average total holding cost across all jobs. However, no optimality results are known for the TVHC problem outside of the asymptotic regime. In this paper, we study a simple yet still challenging special case: A two-class M/M/1 queue in which class 1 jobs incur a non-decreasing, time-varying holding cost and class 2 jobs incur a constant holding cost. Our main contribution is deriving the first optimal (non-decreasing) index policy for this special case of the TVHC problem. Our optimal policy, called LookAhead, stems from the following idea: Rather than considering each job's current holding cost when making scheduling decisions, we should look at their cost some $X$ time into the future, where this $X$ is intuitively called the ``lookahead amount." This paper derives that optimal lookahead amount.

cs.PF↗

arXiv2025

Improving Upon the generalized c-mu rule: a Whittle approach

Scheduling a stream of jobs whose holding cost changes over time is a classic and practical problem. Specifically, each job is associated with a holding cost (penalty), where a job's instantaneous holding cost is some increasing function of its class and current age (the time it has spent in the system since its arrival). The goal is to schedule the jobs to minimize the time-average total holding cost across all jobs. The seminal paper on this problem, by Van Mieghem in 1995, introduced the generalized c-mu rule for scheduling jobs. Since then, this problem has attracted significant interest but remains challenging due to the absence of a finite-dimensional state space formulation. Consequently, subsequent works focus on more tractable versions of this problem. This paper returns to the original problem, deriving a heuristic that empirically improves upon the generalized c-mu rule and all existing heuristics. Our approach is to first translate the holding cost minimization problem to a novel Restless Multi-Armed Bandit (R-MAB) problem with a finite number of arms. Based on our R-MAB, we derive a novel Whittle Index policy, which is both elegant and intuitive.

cs.PF↗

arXiv2024

How to Rent GPUs on a Budget

The explosion in Machine Learning (ML) over the past ten years has led to a dramatic increase in demand for GPUs to train ML models. Because it is prohibitively expensive for most users to build and maintain a large GPU cluster, large cloud providers (Microsoft Azure, Amazon AWS, Google Cloud) have seen explosive growth in demand for renting cloud-based GPUs. In this cloud-computing paradigm, a user must specify their demand for GPUs at every moment in time, and will pay for every GPU-hour they use. ML training jobs are known to be parallelizable to different degrees. Given a stream of ML training jobs, a user typically wants to minimize the mean response time across all jobs. Here, the response time of a job denotes the time from when a job arrives until it is complete. Additionally, the user is constrained by some operating budget. Specifically, in this paper the user is constrained to use no more than $b$ GPUs per hour, over a long-run time average. The question is how to minimize mean response time while meeting the budget constraint. Because training jobs receive a diminishing marginal benefit from running on additional GPUs, allocating too many GPUs to a single training job can dramatically increase the overall cost paid by the user. Hence, an optimal rental policy must balance a tradeoff between training cost and mean response time. This paper derives the optimal rental policy for a stream of training jobs where the jobs have different levels of parallelizability (specified by a speedup function) and different job sizes (amounts of inherent work). We make almost no assumptions about the arrival process and about the job size distribution. Our optimal policy specifies how many GPUs to rent at every moment in time and how to allocate these GPUs.

cs.DC↗

arXiv2022

Graph Searching with Predictions

Consider an agent exploring an unknown graph in search of some goal state. As it walks around the graph, it learns the nodes and their neighbors. The agent only knows where the goal state is when it reaches it. How do we reach this goal while moving only a small distance? This problem seems hopeless, even on trees of bounded degree, unless we give the agent some help. This setting with ''help'' often arises in exploring large search spaces (e.g., huge game trees) where we assume access to some score/quality function for each node, which we use to guide us towards the goal. In our case, we assume the help comes in the form of distance predictions: each node $v$ provides a prediction $f(v)$ of its distance to the goal vertex. Naturally if these predictions are correct, we can reach the goal along a shortest path. What if the predictions are unreliable and some of them are erroneous? Can we get an algorithm whose performance relates to the error of the predictions? In this work, we consider the problem on trees and give deterministic algorithms whose total movement cost is only $O(OPT + Δ\cdot ERR)$, where $OPT$ is the distance from the start to the goal vertex, $Δ$ the maximum degree, and the $ERR$ is the total number of vertices whose predictions are erroneous. We show this guarantee is optimal. We then consider a ''planning'' version of the problem where the graph and predictions are known at the beginning, so the agent can use this global information to devise a search strategy of low cost. For this planning version, we go beyond trees and give an algorithms which gets good performance on (weighted) graphs with bounded doubling dimension.

cs.DS↗

arXiv2022

Analyzing Sharpness along GD Trajectory: Progressive Sharpening and Edge of Stability

Recent findings (e.g., arXiv:2103.00065) demonstrate that modern neural networks trained by full-batch gradient descent typically enter a regime called Edge of Stability (EOS). In this regime, the sharpness, i.e., the maximum Hessian eigenvalue, first increases to the value 2/(step size) (the progressive sharpening phase) and then oscillates around this value (the EOS phase). This paper aims to analyze the GD dynamics and the sharpness along the optimization trajectory. Our analysis naturally divides the GD trajectory into four phases depending on the change of the sharpness. We empirically identify the norm of output layer weight as an interesting indicator of sharpness dynamics. Based on this empirical observation, we attempt to theoretically and empirically explain the dynamics of various key quantities that lead to the change of sharpness in each phase of EOS. Moreover, based on certain assumptions, we provide a theoretical proof of the sharpness behavior in EOS regime in two-layer fully-connected linear neural networks. We also discuss some other empirical findings and the limitation of our theoretical results.

cs.LG↗